Forums

Convergence of adaptive DFE

Started by cpshah99 July 29, 2009
Hi All

I am trying to demodulate real time data using adaptive DFE.

The channel is highly time varying and the multipath spread is approx. 50
- 100 symbols at the data rate of 4ksps.

I have 1000 symbols long training sequence. The problem is that adaptive
DFE is not able to converge. The algorithms that I am using is LMS and
NLMS. I have been palying with filter lengths and step sizes. I can use RLS
but I do not want to do that. Because with 1000 symbols of training any
filter should converge.

My ques: is it true that in highly dynamic channel environment the
adaptive DFE can not track channel variations?

I am going mad because I managed to decode one packet with zero error but
with same setup other packets are not doing any good. 

Your opinion matters a lot.

Chintan

cpshah99 wrote:

> Hi All > > I am trying to demodulate real time data using adaptive DFE. > > The channel is highly time varying and the multipath spread is approx. 50 > - 100 symbols at the data rate of 4ksps.
How "highly" is highly?
> I have 1000 symbols long training sequence. The problem is that adaptive > DFE is not able to converge. The algorithms that I am using is LMS and > NLMS. I have been palying with filter lengths and step sizes.
What is your rationale for selecting the filter length and the step size?
> I can use RLS > but I do not want to do that.
Why? BTW, there are the methods other then RLS.
> Because with 1000 symbols of training any > filter should converge.
How did you decide that?
> My ques: is it true that in highly dynamic channel environment the > adaptive DFE can not track channel variations?
Make the estimates.
> I am going mad because I managed to decode one packet with zero error but > with same setup other packets are not doing any good.
Look for the bugs and the flaws in the processing.
> Your opinion matters a lot.
And what do you expect exactly? Someone should develop the algorithm for you? Stupident :))))) Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
Hello Vlad

Wassup mate????

> >How "highly" is highly?
This is underwater channel and I can see that the multipaths have almost same magnitude.
>What is your rationale for selecting the filter length and the step
size?
>
I got the impulse response by correlating the training seq and from that I decided the number of taps. I guess this is the ideal way.
>> I can use RLS >> but I do not want to do that. > >Why? BTW, there are the methods other then RLS. >
The reason I dont want to use that is complexity reasons. Even though I dont give damn about complexity, I would like to get most out of LMS or NLMS.
>> Because with 1000 symbols of training any >> filter should converge. > >How did you decide that? >
In the past I had used 500 symbols long training which did good job but just to provide good convergence I used 1000 this time.
>> My ques: is it true that in highly dynamic channel environment the >> adaptive DFE can not track channel variations? > >Make the estimates. >
I am going to do that but I also have Doppler effect to take into account. And I dont know to estimate the channel do I need more than one sample per symbol?
>Look for the bugs and the flaws in the processing.
Not really. Because in real time, you will just design your receiver. So it may work or may not. You can't guarantee error free communication always. Out of 20 packets I have managed to decode 15 with same setup but those other 5 packest are creating a pain for me.
>And what do you expect exactly? Someone should develop the algorithm for
>you? Stupident :))))) >
Not exactly. For some reason u always *assume* somebody is asking you to do everything for them. I am not like them. :-) Chintan
On Jul 29, 7:35&#2013266080;am, Vladimir Vassilevsky <nos...@nowhere.com> wrote:
> cpshah99 wrote: > > Hi All > > > I am trying to demodulate real time data using adaptive DFE. > > > The channel is highly time varying and the multipath spread is approx. 50 > > - 100 symbols at the data rate of 4ksps. > > How "highly" is highly? > > > I have 1000 symbols long training sequence. The problem is that adaptive > > DFE is not able to converge. The algorithms that I am using is LMS and > > NLMS. I have been palying with filter lengths and step sizes. > > What is your rationale for selecting the filter length and the step size? > > > I can use RLS > > but I do not want to do that. > > Why? BTW, there are the methods other then RLS. > > > Because with 1000 symbols of training any > > filter should converge. > > How did you decide that? > > > My ques: is it true that in highly dynamic channel environment the > > adaptive DFE can not track channel variations? > > Make the estimates. > > > I am going mad because I managed to decode one packet with zero error but > > with same setup other packets are not doing any good. > > Look for the bugs and the flaws in the processing. > > > Your opinion matters a lot. > > And what do you expect exactly? Someone should develop the algorithm for > you? &#2013266080;Stupident :))))) > > Vladimir Vassilevsky > DSP and Mixed Signal Design Consultanthttp://www.abvolt.com
RLS does not track very well. It's tracking ability relies on an ad- hoc method of a forgetting factor. It is computationally pretty expensive too. Hardy